[RFC] TOTO for zero-shot anomaly detection

[XVmecha]

Dear TOTO team,

I was curious about TOTO's ability to function as an anomaly detection model. I built a zero-shot detection mechanism that calculates the Negative Log-Likelihood of observations under TOTO's predicted Student-t distributions as per-variate error scores. These are aggregated (mean/max) into timestep-level anomaly scores and compared against an unsupervised threshold for binary classification. I made a repository with first results on two Multivariate timeseries anomaly detection benchmark datasets. I'm curious whether you've explored anomaly detection use cases for TOTO, I would be interested to learn about your perspective. The code and a blogpost are avaliable at https://github.com/XVmecha/Toto-4-AD.

All the best,
Andreas Berentzen.